(Sunday, 7/14/2019)
Room: ePoster Forums
Purpose: Lung is one of the most important organs at risk in radiation treatment planning for thoracic patients. Current automatic lung segmentation based on convolutional neural network architecture have achieved success in delineating lung lobes from the thoracic computed tomography scans. However, unfortunately these methods need to train a large number of network parameters thus to require plenty of training data. We present a capsule network model based segmentation strategy to delineate lung on thoracic CT in order to overcome the above shortages.
Methods: The developed deep capsule network architecture based strategy for automatic delineation of lungs consists of two sections, the main segmentation model and the post processing step. The main segmentation model employ 3D UNet as the network architecture but use capsule nodes to replace the original convolutional filters. Since the capsule node groups the scalar feature maps into the feature vectors and routes them in the network through voting strategy, it avoids useless forwarding propagation and focuses the information flows between specific nodes. We use the weighted sum of margin loss, reconstruction loss, dice loss and l1 loss as the loss function to optimize the model. The output of the segmentation model will be refined to generate the final segmentation map through morphology operation such as dilating. The strategy has been validated on the thoracic CT scans dataset provided by 2017 AAPM Thoracic Auto-segmentation. We used the expert annotation Lung_L and Lung_R as the groundtruth.
Results: Experiment results showed the strategy was capable of maintaining the average DICE coefficient(DCs) of 0.80 or 0.93 without or with the post processing steps.
Conclusion: The proposed automatic segmentation strategy could delineate the lung lobes accurately. It will be useful in future clinical practice and helpful in arising the thoracic radiation therapy treatment planning.
Funding Support, Disclosures, and Conflict of Interest: Key research and development plan of Sichuan Province, 2019YFS0125
Keywords
Not Applicable / None Entered.
Taxonomy
Contact Email
